Exjobbspresentation: Beamforming algorithm development for a Neural Processing Unit (NPU)

Exjobbet genomfördes hos BeammWave AB med Samir Kumar Mishra som industrihandledare, Victor Åberg som akademisk handledare och Erik Larsson examinator.

This thesis investigates the development of a digital beamforming algorithm targeting efficient execution on Neural Processing Unit (NPU)-oriented architectures for millimeter-wave (mmWave) Fifth Generation New Radio (5G NR) systems. While mmWave frequencies enable high data rates, they also introduce significant propagation losses that require advanced beamforming techniques that can overcome this limitation reliably.  

Our work focuses on the design and implementation of a precoding algorithm in C, with emphasis on adaptations relevant to NPU deployment, including reduced numerical precision and data-parallel computation. To explore performance characteristics, the algorithm was optimized using Intel's intrinsic instruction sets to leverage Single Instruction Multiple Data (SIMD) execution on a CPU-based platform, serving as a proxy for parallel hardware acceleration.

Results show that reliable and consistent beamforming performance can be maintained under reduced precision, supporting the suitability of such approaches for resource-constrained and parallel architectures. However, the applied SIMD optimizations led to increased execution times, with the Advanced Vector Extension (AVX) based implementation failing to deliver the expected performance improvements.

These findings highlight important tradeoffs between numerical efficiency and computational performance, and emphasize the need for hardware-specific optimization when targeting NPU-like architectures. The work contributes practical insights into the challenges of mapping beamforming algorithms to emerging parallel processing platforms.